All posts
Published — 4 min.

How to go viral on X (part 2)

Sergey Keller

Sergey Keller

GrowthContent
How to go viral on X (part 2) cover image

Three years ago, the public Twitter algorithm discussion focused on older ranking patterns.

The current X stack is a multi-stage system: retrieve candidates, score them with multi-action predictions, apply ranking adjustments, then run hard filters before final delivery.

So if you want to go viral on X now, optimize for what this stack rewards:

  1. High-quality replies, repost/share behavior, and dwell signals.
  2. Low negative feedback (not interested, mute, block, report).
  3. Strong early momentum while the post is fresh.
  4. Content that can perform in-network first, then out-of-network.
  5. Eligibility across all filters (quality/safety/seen/dedup).

What drives virality in the current system

1) Virality is a ranking output, not one metric

X combines in-network and out-of-network candidates, then ranks them with Phoenix predictions and additional scorers.

That means viral performance comes from winning multiple signals together, not one trick.

2) Multi-action engagement is the core score input

The system predicts many actions (favorite, reply, repost, share, dwell, follow, negative actions, etc.) and combines them in a weighted score.

Practical implication: - Optimize for conversation and sharing, not just likes.

3) Negative feedback has explicit downside

Actions like not interested, block, mute, and report are included in scoring with negative effect, so low-trust content can collapse distribution fast.

Practical implication: - Avoid bait formats and misleading hooks.

4) Video is conditional, not automatic upside

Video-quality-view weighting is only applied when the video duration passes a minimum threshold.

Practical implication: - Use video when it improves retention, not by default.

5) Additional ranking layers shape who gets seen

After weighted scoring, author diversity attenuation and out-of-network weighting are applied before top-K selection.

Practical implication: - Fewer, higher-signal posts generally beat repetitive posting.

6) Hard filters can block reach even if score is high

Age limits, seen/served dedup, muted keywords, blocked/muted authors, and visibility filtering can remove posts regardless of score.

Practical implication: - Distribution safety and quality discipline are as important as content quality.

7) Freshness still matters in candidate sourcing

In-network post retrieval is explicitly sorted by recency.

Practical implication: - Early engagement velocity still matters for expansion.

Part 2 tactical checklist

  1. Write a first line that stops scroll and promises a specific payoff.
  2. Include a clear reply trigger (question, stance, challenge).
  3. Make the post easy to share (clear framework, strong one-liner insight).
  4. Minimize negative-feedback risk (no spammy framing, no bait-and-switch).
  5. Show up in the first hour to drive high-quality discussion.
  6. Keep posting cadence high-signal to avoid author fatigue effects.

5 recent viral X examples (February 2026)

Below are 5 recent examples (published in the last days) that align with the ranking principles above.

  1. @feelzyou prompt post (February 9, 2026): “who saved you when you were at your lowest?”
    • Snapshot: ~37,000 likes, ~10,000 replies, ~13M views.
    • Why it fits: extremely high reply intent + emotional relevance (multi-action engagement signal).
    • Source: X trending summary | TwStalker snapshot
  2. @fairiehaze image poll (February 3, 2026): mango vs watermelon
    • Snapshot: ~26,000 likes, ~6.6M views.
    • Why it fits: low-friction interaction format (poll) + high comment/reply potential from binary opinions.
    • Source: X trending summary
  3. @MrMekzy_ “caricature of me and my job” prompt trend (February 6, 2026)
    • Snapshot: ~25,000+ likes on the seed post, then thousands of follow-on reposts/remixes.
    • Why it fits: replication loop (users create their own version), which compounds shares/replies and dwell.
    • Source: X trending summary | TwStalker discovery snapshots
  4. Trader maro “SILVER/USD pendant” post (February 13, 2026)
    • Snapshot: ~37,000 likes, ~1.5M views.
    • Why it fits: timely hook (Valentine’s Day + market move) + visual novelty + high shareability in trader communities.
    • Source: X trending summary
  5. Klara Corvette-owner meme post (February 6, 2026)
    • Snapshot: ~11,000 likes, ~2M views.
    • Why it fits: instantly legible humor + identity-based social sharing + large reply threads with “same here” stories.
    • Source: X trending summary

Important caveat

Some parameter constants (exact weights/thresholds) are intentionally excluded from the open-source release, so we can confirm the scoring structure and action dimensions, but not every production coefficient.

Code Sources (x-algorithm GitHub)

  1. Architecture overview and pipeline stages
  2. Multi-action scoring description
  3. Scorer/filter ordering in Home Mixer pipeline
  4. Phoenix action prediction mapping
  5. Weighted score formula across actions
  6. Video-duration eligibility for VQV weighting
  7. Author diversity attenuation
  8. Out-of-network score adjustment
  9. Age filter (freshness eligibility)
  10. Seen-content filtering
  11. Muted keyword filtering
  12. Blocked/muted author filtering
  13. Visibility filtering (drop actions)
  14. Candidate isolation attention mask in ranking transformer
  15. Ranking model input schema and candidate/history/user features
  16. Action set used by the ranking runner
  17. In-network recency scoring in Thunder
  18. Open-source caveat: params/util modules excluded